AI assistants repeat themselves—it's the temperature setting
The coffee machine hisses in the corner of a Brooklyn co-working space, Monday morning light cutting hard across laptop screens, and somewhere three desks over an operator is pasting the same AI output for the fourth time wondering why ChatGPT suddenly forgot how to vary a sentence.
Temperature is the AI setting that stops you getting copy-paste robots
Set it too low and every response is identical; set it too high and outputs turn incoherent.
Temperature controls randomness in language models. At 0.0, the model picks the most probable next token every time—deterministic, repeatable, and utterly predictable. At 2.0, it samples from the long tail of unlikely tokens, which makes outputs creative in the worst possible way: grammatically intact but semantically unhinged. Most operators never touch the slider, so they live with whatever default their tool ships—usually 0.7 for ChatGPT, 1.0 for Claude, and something undocumented in every no-code AI wrapper.
The symptom shows up fast. You ask for three subject lines and get the same structure with synonyms swapped in. You regenerate a product description and the new version mirrors the old one, word for word, because temperature is locked at zero and the prompt hasn’t changed. The fix is a single slider adjustment, but most platforms bury it three clicks deep or don’t expose it at all. OpenAI’s Playground, Anthropic’s Console, and API-native tools like Promptfoo let you set temperature per request. Zapier’s ChatGPT action, Make’s OpenAI module, and most Notion AI blocks do not.
For operators, the practical range is 0.3 to 0.9. Use 0.3 when you need consistency across a batch—structured data extraction, classified tagging, anything where variance breaks downstream automation. Use 0.7 to 0.9 when you want variety without chaos—email subject lines, social captions, brainstorm lists. Anything above 1.2 is performance art, not production work. If your AI tool doesn’t let you set temperature, you’re renting someone else’s opinion about how random your outputs should be.
RELATED
AI token pricing changes what you can afford to automate
Temperature adjustments are free, but every API call costs tokens—and pricing varies wildly across providers. GPT-4 runs $0.03 per thousand input tokens and $0.06 per output; Claude 3.5 Sonnet charges $0.003 in and $0.015 out. That gap compounds fast when you’re batching hundreds of requests daily. Operators who don’t calculate per-request costs end up shocked by monthly invoices or, worse, quietly throttling workflows that should scale. The math isn’t hidden, but it’s not automatic either.
TACTIC
Track which prompt versions actually worked last month
You’ll tweak temperature, rewrite system instructions, and adjust output length a dozen times before a prompt stabilises—but most operators overwrite the working version and lose it forever. Prompt version control doesn’t require Git or enterprise tooling. A dated Google Doc, a Notion database with timestamps, or a Sheets tab logging prompt text alongside sample outputs will surface patterns fast. When performance drops, you’ll know exactly which version to roll back to instead of guessing what changed between Tuesday and Thursday.
WORTH KNOWING
Context windows reset mid-task and wipe your thread history
Even when temperature is dialled in perfectly, long AI conversations hit a wall: most platforms reset context after 30 to 90 minutes of inactivity. OpenAI’s ChatGPT, Claude’s web interface, and API wrappers all expire sessions differently, and none warn you before the thread resets. If you’re building a prompt iteratively over an afternoon, a lunch break can mean starting from scratch. The workaround is to save intermediate outputs and re-inject context manually, or use tools that persist conversation state across sessions—but you need to know the expiry window exists before you can plan around it.
Know someone who would like this? Forward today’s email—every operator we reach is one closer to running an online business with a little less friction.